fix the deadlock issue when using distributed training in vqa finetune

[Light-V]

The main process with local_rank 0 will wait forever because other process will not call torch.distributed.barrier()